Tencent open-sources its 13B image-to-video AI model

Tencent Hunyuan just open-sourced HunyuanVideo-I2V, releasing the full weights, inference code, and LoRA training scripts to the public.

In plain terms, you feed it a single still image along with a brief prompt describing camera motion or character movement, and it outputs a 5-second clip in crisp 2K resolution, complete with automatically generated background audio.

What it can do

Beyond standard image animation, this release bakes in two practical features: lip-syncing and motion driving.

If you upload a portrait and pair it with text or an audio snippet, the model animates the character's mouth and expressions so they appear to speak or sing naturally.

Motion driving takes it a step further into pose transfer. You can upload a still image of a character and map driving video motions onto it, like generating a full dance sequence. According to technical materials, it works across photorealistic humans, anime figures, and CGI game characters.

If you just want to test it out without setting up an environment, a web demo is up on Tencent's official video platform. Developers looking to host or fine-tune can pull everything directly from GitHub or Hugging Face.

Under the hood and what is included

The image-to-video release builds directly on Hunyuan's text-to-video foundation, keeping the total parameter count at 13 billion.

According to the technical report, both modalities were pretrained on shared datasets. The architecture is built to ingest image, text, audio, and pose inputs simultaneously, giving creators multi-dimensional control over large movements and native camera cuts.

The real win for developers is the included LoRA training code. Rather than wrestling with full-parameter fine-tuning on a 13B model, you can train lightweight adapter weights on your own stylized datasets with reasonable compute budgets.

Tencent's open-source suite now spans text, image, video, and 3D generation, racking up over 23,000 GitHub stars across its repositories.

Why it matters

Open-source video generation has lagged noticeably behind image generation, largely due to massive compute requirements and incomplete tooling releases. Most labs drop weights without clean fine-tuning recipes.

By packaging 13B weights alongside functional LoRA scripts and audio capabilities, Hunyuan provides an actual foundation for indie animators, game developers, and creative technologists. If you have the GPU headroom, this is an immediate candidate for local pipeline testing.

Keep in mind that running and training a 13B video model is not light on VRAM. But in our view, making both the weights and the tuning code freely available puts healthy pressure on proprietary players to offer better controllability and open access.